Table 3. Predictive performance of selected features from 3 subregions and the entire tumor region.

Method	Subregions	Feature	Cohort	AUC	Interval	P†
Intratumoral Regionalization	Rapid	Skewness (S-0)	Training	0.796	0.705–0.888	0.032
			Validation	0.726	0.570–0.882	0.078
		Daubechies2_4_H_V (S-0)	Training	0.548	0.535–0.603	0.021
			Validation	0.515	0.315–0.715	0.813
		SRHGLE_0 (S-4)	Training	0.741	0.646–0.836	0.128
			Validation	0.584	0.397–0.771	0.599
		GLN_135 (S-4)	Training	0.628	0.514–0.743	0.969
			Validation	0.621	0.434–0.807	0.417
		Skewness (S-8)	Training	0.805	0.719–0.892	0.011
			Validation	0.737	0.581–0.893	0.085
	Medium	Daubechies2_4_H_V (S-4)	Training	0.724	0.624–0.825	0.263
			Validation	0.640	0.468–0.812	0.347
	Slow	Skewness (S-0)	Training	0.651	0.538–0.764	0.816
			Validation	0.608	0.417–0.799	0.546
		Skewness (S-4)	Training	0.713	0.613–0.813	0.299
			Validation	0.616	0.436–0.797	0.458
		RP_0 (S-4)	Training	0.604	0.493–0.716	0.745
			Validation	0.625	0.450–0.800	0.414
		Harr_4_H_H (S-4)	Training	0.639	0.528–0.751	0.930
			Validation	0.621	0.437–0.805	0.444
		Harr_4_H_D (S-4)	Training	0.742	0.638–0.846	0.142
			Validation	0.603	0.436–0.771	0.451
Whole Lesion	–	Covariance (S-8)	Training	0.632	0.520–0.744	–
			Validation	0.522	0.343–0.701	–

^†, P value represents the performance comparison between the features from the intratumoral subregions and the feature from the whole tumor region. AUC, area under the receiver operating characteristic curve; SRHGLE, short-run high gray-level emphasis; GLN, gray-level nonuniformity; RP, run percentage.